Quantum automorphism groups of homogeneous graphs

Transcription

1 Journal of Functional Analysis 224 (2005) wwwelseviercom/locate/jfa Quantum automorphism groups of homogeneous graphs Teodor Banica Department of Mathematics, Universite Paul Sabatier, 118 route de Narbonne, Toulouse, France Received 25 June 2004; received in revised form 17 November 2004; accepted 18 November 2004 Communicated by Dan Voiculescu Available online 4 January 2005 Abstract Associated to a finite graph X is its quantum automorphism group G The main problem is to compute the Poincaré series of G, meaning the series f(z)= 1 + c 1 z + c 2 z 2 + whose coefficients are multiplicities of 1 into tensor powers of the fundamental representation In this paper we find a duality between certain quantum groups and planar algebras, which leads to a planar algebra formulation of the problem Together with some other results, this gives f for all homogeneous graphs having 8 vertices or less 2005 Elsevier Inc All rights reserved Keywords: Quantum permutation group; Planar algebra; Fuss Catalan algebra; Tannakian duality 0 Introduction A remarkable discovery, due to Wang [14], is that the set X n ={1,,n} has a quantum automorphism group, bigger in general than the symmetric group S n The quantum group does not exist of course, but the algebra of continuous functions on it does This is a certain Hopf C -algebra constructed with generators and relations, denoted here H(X n ) For n = 1, 2, 3 the quotient map H(X n ) C(S n ) is an isomorphism For n 4 it is not, and in fact H(X n ) is non-commutative, and infinite dimensional address: banica@picardups-tlsefr (T Banica) /$ - see front matter 2005 Elsevier Inc All rights reserved doi:101016/jjfa

2 244 T Banica / Journal of Functional Analysis 224 (2005) There are several variations of this construction, see for instance [5,3] The idea is that Hopf algebra quotients of H(X n ) correspond to quantum automorphism groups of various discrete objects, like finite graphs, finite metric spaces, and so on A first purpose of this paper is to formulate some precise problem concerning such universal Hopf algebras We believe that the following statement is the good one Problem 01 Let X be a finite graph, all whose edges are colored and possibly oriented, such that an oriented edge and a non-oriented one cannot have same color We denote by H(X) the universal Hopf C -algebra coacting on X This is obtained as an appropriate quotient of Wang s algebra H(X n ), where n is the number of vertices of X The problem is to compute its Poincaré series f(z)= 1 + c 1 z + c 2 z 2 +, where c k is the multiplicity of 1 into the kth tensor power of the fundamental coaction of H(X) As a first remark, in the computation X f both the input and the output are classical objects The problem is to find a classical computation relating them The input is a bit more general than in our previous paper [3] Indeed, the algebra associated in [3] to a finite metric space Y is the same as the algebra H(Y gr ), where Y gr is the complete graph having vertices at points of Y, with edges colored by corresponding lengths As for the output, the choice of f instead of other invariants, like fusion rules, etc is inspired from recent progress in subfactors, such as Jones s fundamental work [11] There is actually one more subtlety here, concerning the input The interesting case is when X is quantum homogeneous, meaning that the algebra of fixed points of the universal coaction reduces to the scalars (This is the same as asking for the equality c 1 = 1, known to correspond to the irreducibility condition in subfactors) It follows from definitions that if X is homogeneous then it is quantum homogeneous The converse appears to be true in many cases of interest, but so far we do not know if it is true in general The main tool for solving problem 01 is Woronowicz s Tannakian duality [16] In this paper, we find a general result in this sense This is a duality between Hopf algebra quotients of H(X n ) and subalgebras of Jones s spin planar algebra [9], which preserves Poincaré series The algebra H(X) corresponds in this way to the planar algebra generated by the incidency matrices of X, one for each color, viewed as 2-boxes in the spin planar algebra The main application is with a product of s complete graphs The s incidency matrices satisfy Landau s exchange relations [13], so we get a Fuss Catalan algebra on s colors, whose Poincaré series is computed by Bisch and Jones in [6] The s = 2 case of this result was previously obtained in [3], as a corollary of a direct combinatorial computation in [2] We get in this way a Fuss Catalan series of graphs, containing products of complete graphs, plus other graphs, obtained by performing modifications which preserve the H algebra Another series is the dihedral one, where we have H(X) = C(D n ) It is known from [3] that n-gons with n = 4 belong to it, and here we find an improved statement

3 T Banica / Journal of Functional Analysis 224 (2005) It is known from [3] that the Fuss Catalan and dihedral series cover all non-oriented homogeneous graphs having n 7 vertices In this paper we prove the following result Theorem 01 The non-colored non-oriented homogeneous graphs having n 8 vertices fall into three classes: (1) Fuss Catalan graphs, (2) dihedral graphs, (3) the cube and its complement, which correspond to a tensor product between TL(2) and TL(4) The case of colored or oriented graphs is discussed as well 1 Formalism Let H be a C -algebra with unit, together with C -morphisms Δ : H H H, ε : H C and S : H H op called comultiplication, counit and antipode Here is any C -algebra tensor product and H op is the C -algebra H, but with opposite product We assume that the square of the antipode is the identity, and that Woronowicz s axioms in [17] are satisfied Let X be a finite set We denote by C(X) the algebra of complex functions on X The linear form on C(X) which sums the values of the function is denoted Σ Definition 11 A coaction of H on X is a morphism of C -algebras v : C(X) C(X) H satisfying the following conditions (a) Coassociativity condition (id Δ)v = (v id)v (b) Counitality condition (id ε)v = id (c) Natural condition (Σ id)v = Σ()1 We should mention that in this definition the terminology is not standard The natural condition says that the action of the corresponding quantum group must preserve the counting measure on X This condition is satisfied in all reasonable situations, but is not automatic in general See [14] Consider the basis of C(X) formed by Dirac masses Any linear map v : C(X) C(X) H can be written in terms this basis, and the matrix of coefficients (v ij ) determines v v(δ i ) = j δ j v ji The simplest example involves the Hopf C -algebra C(G) associated to a finite group G Here the comultiplication, counit and antipode are obtained by applying the C functor to the multiplication, unit and inverse map of G If G acts by permutations on X, we can apply the C functor to the corresponding map (x, g) g(x) and we get a coaction of C(G) on X

4 246 T Banica / Journal of Functional Analysis 224 (2005) Theorem 11 Let G be a group of permutations of a finite set X The coefficients v ij of the corresponding coaction of C(G) on X are characteristic functions of the sets {g G g(j) = i} When i is fixed and j varies, or vice versa, these sets form a partition of G Back to the general case, it is convenient to translate the conditions in Definition 11 in terms of coefficients v ij First, the Dirac masses being self-adjoint, the fact that v is a -map says that the elements v ij are self-adjoint Multiplicativity of v says that v(δ i )v(δ k ) = v(δ i δ k ) for any i, j, k This translates into the following formulae, where δ ik is a Kronecker symbol δ j v ji v jk = δ ik δ j v ji j For i = k we must have that v ji is a projection, and for i = k we must have that v ji is orthogonal to v jk The fact that v is unital translates into the following formula: δ j v ji = δ j 1 ij j Summing up, the fact that the linear map v is a morphism of C -algebras is equivalent to the fact that all rows of the matrix v are partitions of unity with self-adjoint projections The natural condition says that the sum on each column of v is 1 Thus v is a unitary corepresentation, so the antipode is given by S(v ij ) = v ji By applying the antipode we get that the columns of v are also partitions of unity In other words, the linear map produced by a matrix v is a coaction if and only if the coassociativity, counitality and natural conditions are satisfied, and v is a magic biunitary in the following sense Definition 12 A matrix v M X (H ) is called magic biunitary if its rows and columns are partitions of unity of H with self-adjoint projections A magic biunitary v is indeed a biunitary, in the sense that both v and its transpose v t are unitary matrices The terminology comes from a vague similarity with magic squares A coaction is said to be faithful if its coefficients generate the C -algebra H We are interested in faithful coactions, and it is convenient to translate both axioms for H and for v in terms of coefficients, by formulating the above discussion in the following way Theorem 12 If v is a faithful coaction of H on X the following conditions are satisfied (a) The matrix v = (v ij ) is a magic biunitary and its coefficients generate H j

5 T Banica / Journal of Functional Analysis 224 (2005) (b) There is a C -morphism Δ : H H H such that Δ(v ij ) = v ik v kj (c) There is a C -morphism ε : H C such that ε(v ij ) = δ ij (d) There is a C -morphism S : H H op such that S(v ij ) = v ji Conversely, a C -algebra H together with a matrix v M X (H ) satisfying these conditions has a unique Hopf C -algebra structure such that v is a faithful coaction ofhonx This statement is similar to Definition 11 in Woronowicz s paper [15] The objects (H, v) satisfying the above conditions correspond to compact permutation pseudogroups Woronowicz s analogue of the Peter Weyl theorem in [15] shows that each irreducible corepresentation of H appears in a tensor power of v The main problem is to decompose these tensor powers, given by the following formulae: v k (δ i1 δ ik ) = (δ j1 δ jk ) v j1 i 1 v jk i k By Frobenius reciprocity an equivalent problem is to compute spaces of fixed points Hom(1,v k ) ={x C(X) k v k (x) = x 1} A slightly easier problem is to compute dimensions, arranged in a Poincaré type series Definition 13 The Poincaré series of a coaction v : C(X) C(X) H is f(z)= k=0 ( ) dim Hom(1,v k ) z k, where Hom(1,v k ) is the space of fixed points of the kth tensor power of v Woronowicz s results in [15] show that the dimension of the fixed point space of a corepresentation is obtained by applying the Haar functional h : H C to its character By using the multiplicativity of the character map χ we get the following equality: ( ) ( ) dim Hom(1,v k ) = h χ k v The diagonal entries of v are self-adjoint projections, so their sum χ v is self-adjoint For z small enough the operator 1 zχ v is invertible, with inverse given by the following

6 248 T Banica / Journal of Functional Analysis 224 (2005) formula: (1 zχ v ) 1 = z k χ k v By the above, applying the Haar functional gives the Poincaré series Theorem 13 The Poincaré series of a coaction v : C(X) C(X) H is given by k=0 ( f(z)= h 1 z i ) 1 v ii, where h : H C is the Haar functional If v is faithful the convergence radius of f is bigger than 1/(#X), with equality if and only if H is amenable in the Hopf C -algebra sense In this statement the first assertion follows from the above discussion The second one follows from a quantum analogue of the Kesten amenability criterion for discrete groups, due to Skandalis, and written down in the last section of [1] For H = C(G) the Poincaré series is given by the following formula, where π is the corresponding representation of G on C(X), and where χ π is its character f(z)= 1 1 G 1 χ g G π (g)z The number χ π (g) is the trace of g, viewed as an operator from C(X) to itself The Hilbert space structure of C(X) is the one making the Dirac masses an orthonormal basis Since g permutes elements of the basis, its trace χ π (g) is the number m g of its fixed points Theorem 14 Let G be a group of permutations of a set X with n elements The Poincaré series of the corresponding coaction of C(G) is given by the formula f(z)= 1 G n m=0 #G m 1 mz, where G m G is the subset of permutations having exactly m fixed points The convergence radius is 1/n Trying to compute Poincaré series will be the main objective in this paper For a non-commutative Hopf C -algebra H this is in general a quite complicated analytic

7 T Banica / Journal of Functional Analysis 224 (2005) function, and does not have such a simple decomposition as a sum For instance in the non-amenable case computation of the convergence radius is known to be a delicate problem 2 Coactions on abstract spaces Let X be a finite set and let d = (d ij ) be a complex matrix with indices in X If a group G acts on X as in Section 1 then both d and v = (v ij ) are matrices with coefficients in H and indices in X, so we can form the products dv and vd Theorem 21 Let G be a group of permutations of a finite set X and let d M X (C) be a complex matrix with indices in X Consider the magic biunitary v describing the corresponding coaction of C(G) The action of G preserves coefficients of d if and only if dv = vd Proof The product dv is computed by using the formula for v ij in Section 1 (dv) ij = k d ik v kj = k g(j)=k d ik δ g = g d ig(j) δ g The same method gives a similar formula for the product vd (vd) ij = k v ik d kj = k g(k)=i d kj δ g = g d g 1 (i)j δ g With i = g(l) we get that d lj = d g(l)g(j) holds for any l,j,g if and only if dv = vd Example 21 Finite metric spaces Let X be a finite metric space That is, we are given a finite set X and a real function d : X X R which is zero on the diagonal, positive outside, and whose values satisfy the triangle inequality The distance function d can be regarded as a complex matrix with indices in X Then the action of G is isometric if and only if dv = vd This follows from Theorem 21 Indeed, the action is isometric when d ij = d g(i)g(j) for any i, j X, and this means that G preserves coefficients of d in the sense of Theorem 21 Example 22 Finite graphs Let X be a finite graph That is, we are given a finite set X, whose elements are called vertices, and edges are drawn between certain pairs of different vertices The edges are uniquely determined by the incidency matrix, given by d ij = 1ifi, j are

8 250 T Banica / Journal of Functional Analysis 224 (2005) connected by an edge and d ij = 0 if not Then the action of G preserves the edges if and only if dv = vd This is another application of Theorem 21 For, recall first that G preserves the edges of a graph with vertex set X when (ij) is an edge (g(i)g(j)) is an edge In terms of the incidency matrix, this condition is d ij = 1 d g(i)g(j) = 1 Now since d isa0 1 matrix this is the same as asking for the equalities d ij = d g(i)g(j) for any i, j X, so Theorem 21 applies Example 23 Finite oriented graphs Let X be a finite oriented graph That is, X is a graph all whose edges have an orientation We associate to X matrices r and s as follows If ij is an oriented edge we set r ij = 1, r ji = 0 and s ij = i, s ji = i These matrices are related by the formula s = ir ir t Each of them determines X The matrix r has the advantage of being real and the matrix s has the advantage of being self-adjoint The conditions rv = vr and sv = vs are equivalent, and are satisfied if and only if the action of G preserves the oriented edges This is seen as follows First, the action of G preserves the edges of an oriented graph with vertex set X when (ij) is an oriented edge (g(i)g(j)) is an oriented edge In terms of r this is equivalent to r ij = 1 r g(i)g(j) = 1, and since r isa0 1 matrix this is the same as asking for the equalities r ij = r g(i)g(j) for any i, j X Thus Theorem 21 applies and shows that G preserves the oriented edges if and only if rv = vr In terms of s we have that G preserves the oriented edges if and only if s ij =±i s g(i)g(j) =±i, for any i, j X and for any choice of the sign ± Since ±i are the only non-zero coefficients of s, this is the same as asking for s ij = s g(i)g(j) for any i, j X, so Theorem 21 applies In general, commutation of v with d does not really depend on the precise value of coefficients d ij What matters is whether various pairs of coefficients d ij and d kl are equal or not The same happens for an arbitrary coaction v and an arbitrary matrix d Theorem 22 Let v be a coaction on a finite set X and let d M X (C) be a matrix Consider the decomposition d = cd c, where for c C the matrix d c is defined by (d c ) ij = 1 if d ij = c and (d c ) ij = 0 if not, and where the sum is over non-zero terms Then (v ij ) commutes with d if and only if it commutes with all matrices d c Proof We follow the proof in [3] The magic biunitarity condition shows that the multiplication M : δ i δ j δ i δ j and comultiplication C : δ i δ i δ i intertwine v 2 and v Their iterations M (k) and C (k) intertwine v k and v, so the following operator commutes with v d (k) = M (k) d k C (k) = c C c k d c

9 T Banica / Journal of Functional Analysis 224 (2005) Let S be the set of complex numbers c such that d c = 0 Consider the function f : S C given by c c for any c This function separates points of S, and by the Stone Weierstrass theorem the subalgebra of C(S) generated by f must be C(S) itself In particular for any e S the Dirac mass at e is a linear combination of powers of f δ e = λ k f k = ( ) λ k c k δ c = ( ) λ k c k δ c k k c S c S k The corresponding linear combination of matrices d (k) is given by the following formula λ k d (k) = k k ( ) λ k c k d c = ( ) λ k c k d c c S c S k Dirac masses being linearly independent, in the first formula all coefficients in the right term are 0, except for the coefficient of δ e, which is 1 Thus the right term in the second formula is d e It follows that d e is the algebra End(v) of operators commuting with v Another useful decomposition of d is the spectral decomposition, in case d is selfadjoint If so, the C -algebra generated by d is spanned by spectral projections, and commutation with v gives invariant subspaces, by using the following simple fact Theorem 23 Let v be a coaction and let K be a linear subspace of C(X) The matrix (v ij ) commutes with the projection onto K if and only if v(k) K H Proof Let P be the projection onto K For any point k X we have the following formula ( ) v(p(δ k )) = v P ik δ i = P ik δ j v ji = i ij j δ j i v ji P ik On the other hand the linear map (P id)v is given by a similar formula (P id)(v(δ k )) = i P(δ i ) v ik = ij δ j P ji v ik = j δ j i P ji v ik It follows that vp = (P id)v is equivalent to the following conditions, for any j,k: v ji P ik = i i P ji v ik

10 252 T Banica / Journal of Functional Analysis 224 (2005) In other words, the equality of linear maps vp = (P id)v is equivalent to the equality of products of square matrices vp = Pv, and the conclusion follows The universal Hopf C -algebra coacting on X is constructed by Wang in [14] Its quotient by the relations vd = dv is a universal object for the notion of coaction we are interested in Theorem 24 Let X be a finite set and let d M X (C) be a complex matrix Consider the universal C -algebra H(X,d) defined with generators v ij with i, j X and with the relations making v = (v ij ) a magic biunitary matrix commuting with d Then H(X,d) is a Hopf C -algebra and v is a coaction of it on X Proof The universal C -algebra exists because its generators v ij are projections, whose norms are bounded by 1 Call it H Consider the following matrix with coefficients in H H : w ij = k v ik v kj Since v is a magic biunitary, w is magic biunitary as well We multiply to the right by d (wd) ij = k v ik (vd) kj = k v ik (dv) kj = kl d kl v ik v lj On the other hand, the product dw is given by the same formula (dw) ij = l (dv) il v lj = l (vd) il v lj = kl d kl v ik v lj Thus w is a magic biunitary commuting with d, and the formula Δ(v ij ) = v ik v kj defines a C -morphism Also, the identity matrix is a magic biunitary commuting with d, so the formula ε(v ij ) = δ ij defines a C -morphism Consider now the transpose matrix v t, whose coefficients are viewed as elements of the opposite algebra H op Then v t is a magic biunitary, and commutation with d follows from the following computation in H: v t d = v d = v d(vv ) = v (dv)v = v (vd)v = (v v)dv = dv = dv t Thus S(v ij ) = v ji defines a C -morphism, and all conditions in Theorem 12 are satisfied In order to cut off unwanted complexity, the very first condition to be put on (X, d) is quantum homogeneity This is a condition which goes under various names in [3] it

11 T Banica / Journal of Functional Analysis 224 (2005) is called quantum transitivity stating that the algebra of functions fixed by v reduces to C This is the same as saying that the associated planar algebra or subfactor is irreducible Definition 21 We say that (X, d) is quantum homogeneous if v(f ) = f id implies that f is a constant function, where v is the universal coaction of H(X,d) The main example is when (X, d) is homogeneous, meaning that its symmetry group G acts transitively (That is, for any i, j X there is a permutation σ : X X which preserves d, such that σ(i) = j) Indeed, transitivity of G is equivalent to the fact that w(f ) = f id implies that f is constant, where w is the corresponding coaction of C(G), and by using the universal property of v we get that (X, d) is quantum homogeneous We do not know if the converse holds, namely if quantum homogeneous implies homogeneous, but we have the following useful criterion here Theorem 25 Let d M X (C) be a complex matrix For any complex number c the characteristic function of {i d ii = c} is fixed by the universal coaction of H(X,d) In particular if (X, d) is quantum homogeneous all diagonal entries of d must be equal Proof Let f be the characteristic function By using Theorem 22 we may assume that d is a 0 1 matrix and that c = 1 Consider the decomposition of the 0 1 ( 1) matrix d 1 d 1 = (d 1) 1 (d 1) 1 Theorem 22 shows that v commutes with P = (d 1) 1 Since 1 values can appear only on the diagonal of d 1, the 0 1 matrix P is diagonal and 1 P is the projection onto Cf Thus v commutes with the projection onto Cf and Theorem 23 applies v(cf) (Cf) H(X,d) We can write v(f ) = f a From the natural condition we get Σ(f )a = Σ(f )1, so a = 1 3 Colored semi-oriented graphs We are interested in spaces (X, d) which are quantum homogeneous Theorem 25 shows that all diagonal entries of d must be equal By substracting a scalar multiple of the identity we may assume that d is zero on the diagonal This framework is still too general, and we do not have further results at this level For the rest of the paper we make the quite natural assumption that d is self-adjoint

12 254 T Banica / Journal of Functional Analysis 224 (2005) So, assume that d is self-adjoint and has 0 on the diagonal We call vertices the elements of X For any i and j consider the complex number c = d ij = d ji Ifc = 0 we do nothing, if c is real we draw the edge ij and color it c, if the imaginary part of c is positive we draw the oriented edge ij and color it c, and if the imaginary part of c is negative we draw the oriented edge ji and color it c We get a picture, that we call colored semi-oriented graph Definition 31 A colored semi-oriented graph X is a finite graph with all edges colored and possibly oriented, such that an oriented edge and a non-oriented one cannot have same color The choice of colors is not part of X For each color c consider the semi-oriented graph X c obtained by removing all edges having color different from c, then by considering that remaining edges are no longer colored c These are graphs and oriented graphs, called color components of X The incidency matrices of a graph and of an oriented graph are by definition the matrices d and s in Examples 22 and 23 The incidency matrices of colored components of X are called incidency matrices of X and are denoted d c Those corresponding to graphs are 0 1 matrices, those corresponding to oriented graphs are 0 i ( i) matrices They are all self-adjoint The collection of all incidency matrices determines X Definition 32 Associated to X is the universal Hopf C -algebra H(X) coacting on the set of vertices, such that the matrix of coefficients (v ij ) commutes with all incidency matrices of X Both existence and uniqueness follow from Theorems 22 and 24 Note that for graphs our notion of coaction is different from Bichon s notion [5], where H has a coaction on the algebra of functions on the set of edges, compatible with v in some natural sense In the H = C(G) case both our notion of coaction and Bichon s coincide with the usual notion for groups In the general case they are different For instance when X is the complete graph with 4 vertices Definition 32 gives Wang s algebra in [14], which is infinite dimensional, while Bichon s universal construction produces the algebra C(S 4 ), cf comments before Proposition 33 in [5] Let X and Y be colored semi-oriented graphs having the same vertex set By analogy with usual symmetry groups, we say that H(X) is bigger than H(Y) if there exists a Hopf C -algebra morphism H(X) H(Y) mapping coefficients of the universal coaction on X to corresponding coefficients of the universal coaction on Y This happens precisely when for any coaction v of a Hopf C -algebra on the vertex set we have that commutation of (v ij ) with the incidency matrices of Y implies commutation of (v ij ) with the incidency matrices of X It follows from definitions that bigger and smaller imply equal Note also that the algebra of functions on the symmetry group G(X) is smaller than H(X) Theorem 31 Let X be a colored semi-oriented graph (a) Removing a color component increases H(X)

13 T Banica / Journal of Functional Analysis 224 (2005) (b) Reversing orientation in a color component does not change H(X) (c) Forgetting orientation in a color component increases H(X) (d) Identifying two different colors, assumed to color same type of edges, increases H(X) (e) Adding all missing edges, unoriented and colored with a new color, does not change H(X) Proof Removing a color component means removing a commutation relation in definition of H, and we get (a) At level of incidency matrices reversing orientation is given by d d, and we get (b) It is enough to prove (c) for an oriented graph Here the incidency matrix has 0 1 decomposition of type s = is i is i, and the new incidency matrix is given by d = s i + s i Ifv commutes with s it must commute with both matrices s ±i, so it commutes with d as well It is enough to prove (d) for graphs, oriented or not In both cases commutation of v with two matrices is replaced by commutation with their sum or difference, and we get (d) From (a) we get that adding all missing edges decreases H(X) Thus in (e) it is enough to prove that H(X) increases For, let X b be obtained from X by forgetting all orientations and identifying all colors, say with a black color The incidency matrices of X b and of the new color component, say a white color component, are related by the following formula: d b + d w = Let v be the universal coaction of H(X)IfI denotes the matrix filled with 1 then left and right multiplication by I is making sums on rows and columns, and the magic biunitary condition shows that both vi and Iv are equal to I Thus v commutes with the matrix on the right On the other hand we know from (c) and (d) that H(X b ) is bigger than H(X),sov commutes with d b It follows that v commutes with d w and we are done Theorem 32 If X is a graph and X c is its complement then H(X) = H(X c ) Proof From (e) we get that both H(X) and H(X c ) are equal to H of the complete graph having color components X and X c It is convenient to give names to pictures Most geometric objects are metric spaces, and in the non-oriented case we can use (e) plus the following consequence of Theorem 22 Theorem 33 To any finite metric space we associate the colored graph X obtained by drawing edges between all pairs of points and coloring them with their lengths Then H(X) is isomorphic to the universal Hopf C -algebra coacting on the space in

14 256 T Banica / Journal of Functional Analysis 224 (2005) a co-isometric way, in the sense that matrix of coefficients (v ij ) commutes with the distance matrix In general, when drawing a picture of a geometric object what happens is that the symmetry group of the picture, regarded as a graph, is equal to the symmetry group of the rigid object, regarded as a metric space The same is true for associated Hopf algebras Example 31 Simplex vs simplex Consider the n-simplex, viewed as metric space with n points Consider also the usual picture of the n-simplex, viewed as a graph with n vertices and ( n 2 ) edges The distance matrix of the metric space is proportional to the incidency matrix of the graph Thus H of the n-simplex metric space is equal to H of the n-simplex graph Example 32 Cube vs cube Consider the cube, viewed as metric space with 8 points Consider also the usual picture of a cube, viewed as a graph with 8 vertices and 12 edges The 0 1 decomposition of the distance matrix of the metric space is as follows, where a is the length of the side d = ad a + 2ad 2a + 3ad 3a The 0 1 matrix d a is the incidency matrix of the graph Its square counts 2-loops on the graph, and we get da 2 = 3 + 2d 2a Thus a coaction v on the set of 8 vertices commuting with d 1 must commutes with d 2a, and from Theorem 31 we get that v commutes with d 3a as well Thus v must commute with d Conversely, if v commutes with d then Theorem 22 shows that v commutes with d a Commutation with d is equivalent to commutation with d a Thus H of the cube metric space is equal to H of the cube graph Example 33 Polygon vs polygon Consider the regular n-gon, viewed as a metric space Consider also the usual picture of the regular n-gon, viewed as a graph The 0 1 decomposition of the distance matrix of the metric space is as follows, where a is the length of the sides and b<c< are the lengths of various diagonals d = ad a + bd b + cd c + The 0 1 matrix d a is the incidency matrix of the graph Counting 2-loops gives da 2 = 2 + d b, so commutation with d a is equivalent to commutation with both d a and d b The picture shows that d a d b = d a +d c, d a d c = d b +d d and so on, and by induction

15 T Banica / Journal of Functional Analysis 224 (2005) we get that commutation with d a is equivalent to commutation with all 0 1 components of d Thus H of the regular n-gon metric space is equal to H of the regular n-gon graph Theorem 34 Let X be a graph which is quantum homogeneous, in the sense that v(f ) = f 1 implies that f is constant, where v is the universal coaction commuting with the incidency matrix For any l 2 the number of l-loops based at a vertex is independent of the vertex Proof This follows from the fact that the diagonal entry d l (p, p) of the lth power of the incidency matrix counts l-loops at p, and in Theorem 25 one can replace d by any of its powers 4 Cyclic and dihedral groups The simplest example of oriented graph is the oriented n-gon This graph has vertices 1,,n and an oriented edge joins i and i + 1 for any i, with i taken modulo n Theorem 41 The Hopf C -algebra associated to the oriented n-gon is the algebra of functions on the cyclic group Z n The Poincaré series is given by and its convergence radius is 1/n f(z)= 1 + z 1 nz Proof The oriented n-gon X has the following real incidency matrix: r = This is a permutation matrix If v is the coaction of H(X), commutation of v with r says that v must be of the following special form: v 1 v 2 v 3 v n v n v 1 v 2 v n 1 v = v 3 v 4 v 5 v 2 v 2 v 3 v 4 v 1

16 258 T Banica / Journal of Functional Analysis 224 (2005) Since v is a magic biunitary, the elements v i form a partition of the unity of H(X) In particular they commute, so H(X) is commutative Thus H(X) is the algebra of functions on the usual symmetry group of X, which is the cyclic group Z n The Poincaré series is computed by using Theorem 14 The unit of Z n has n fixed points, and the other n 1 elements, none f(z)= 1 n ( n ) = ( ) 1 nz n 1 nz = nz n 1 nz This is equal to the function in the statement In the non-oriented case, the Hopf C -algebra associated to the n-gon is infinite dimensional if n = 4, and equal to the algebra of functions on the dihedral group D n if n = 4 For n 4 this is known from Wang s paper [14] and for n 5 this is proved in [3] We extend now the n = 4 result to a larger class of cyclic graphs Definition 41 We say that a graph X is cyclic if its automorphism group contains a copy of the cyclic group Z n, where n is the number of vertices If X is cyclic, one can choose a vertex and label it 0, then label the other vertices 1,,n 1 such that Z n acts by g(k) = g + k, with both g and k modulo n It is useful to keep in mind the following interpretation Vertices of X are nth roots of unity in the complex plane, counted counterclockwise starting with 1, and edges are segments joining vertices The graph is cyclic if the 2π/n rotation of the plane leaves invariant the picture Definition 42 If X is a cyclic graph with vertices labeled 0, 1,,n 1 we define numbers e(k) by e(k) = 1if0andk are connected by an edge and e(k) = 0 if not Then Q(z) = e(1)z + e(2)z 2 + +e(n 2)z n 2 + e(n 1)z n 1 is a polynomial which does not depend on the choice of the vertex 0 This is a sum of monomials which is symmetric with respect to z n/2, in the sense that Q has degree at most n 1 and the coefficient of z k is equal to the coefficient of z n k, for any k Any sum of monomials which is symmetric with respect to z n/2 is of this form The simplest cyclic graph is the n-gon, corresponding to Q(z) = z + z n 1 Theorem 42 Let X be a cyclic graph with n = 4 vertices and consider the associated polynomial Q Let w be a primitive nth root of unity If the numbers Q(1), Q(w), Q(w 2 ),,Q(w [n/2] )

17 T Banica / Journal of Functional Analysis 224 (2005) are distinct then H(X) is the algebra of functions on D n The Poincaré series is f(z)= 1 + z 2 ( 1 1 εz + 1 ), 1 nz where ε = 1 if n is odd and ε = 2 if n is even The convergence radius is 1/n This is an extension of the result for the n-gon Indeed, for the n-gon having vertices at roots of unity 1,w,w 2,,w n 1 the number Q(w k ) = w k +w n k for w k above the x-axis is twice the projection of w k on the x-axis, which decreases when k increases Proof This follows proofs in [14,3] We use the Vandermonde formula 0 e(1) e(2) e(n 1) e(n 1) 0 e(1) e(n 2) e(1) e(2) e(3) 0 1 w k w 2k w (n 1)k = Q(wk ) 1 w k w 2k w (n 1)k The eigenvalues Q(w k ) being distinct, we get the list of invariant subspaces of d C1, Cξ Cξ n 1, Cξ 2 Cξ n 2, Here ξ = (w i ) and the last subspace has dimension 1 or 2, depending on the parity of n We claim that v is given by formulae F k of following type, with k = 1, 2,,n 1 v(ξ k ) = ξ k a k + ξ n k b k This follows by applying many times Theorem 23 For n = 2 we take b = 0, we define a by F 1, and F k follows by induction For n 3 we define a,b by F 1 Taking the square and cube of F 1 gives ab = ba and ab 2 = ba 2 = 0 With these relations, F k follows by induction Applying to F 1 and comparing with F n 1 gives a = a n 1 and b = b n 1 Together with ab 2 = 0 this gives abb a = 0 Thus ab = ba = 0, and in particular H(X) is commutative On the other hand H(X) depends only on spectral projections of d, so it must be the same for all graphs in the statement With the n-gon we get H(X) = C(D n ) For n = 1 the Poincaré series is computed by using Theorem 14 f(z)= 1 1 z = 1 + z 1 1 z = 1 + z ( z + 1 ) 1 z

18 260 T Banica / Journal of Functional Analysis 224 (2005) For n = 2 the group G has two elements The identity is in G 2 and the other element is in G 0 This gives the formula in the statement f(z)= 1 2 ( ) = 1 z 1 2z 1 2z = 1 + z 1 2z = 1 + z ( z + 1 ) 1 2z For n 3 odd the group G has 2n elements The identity is in G n, the n 1 rotations are in G 0 and the n symmetries are in G 1 Theorem 14 applies and gives the result f(z)= 1 ( n 1 + n 2n 1 z + 1 ) 1 nz (( ) ( )) n 1 = n = n 1 z n + ( nz 1 z + nz 1 nz ) 1 nz 1 For n 4 even the group G has 2n elements as well The identity is in G n and the n 1 rotations are in G 0 There are n more elements, namely the symmetries, half of them being in G 0 and half of them being in G 2 We apply Theorem 14 f(z)= 1 ( 3n/2 1 + n/2 2n (( n/2 = n = n 1 2z + 1 ) 1 nz ) 1 2z n/2 + ( nz 1 2z + nz ) 1 nz ( 1 1 nz 1 )) The proof of Theorem 42 is now complete We apply now Theorem 42 to graphs with small number of vertices We call first and second 9-star the graphs corresponding to the following polynomials, with e = 1, 2 Q(z) = z + z 1+e + z 8 e + z 8 The 8-spoke wheel is the graph corresponding to Q(z) = z + z 4 + z 7 Corollary 41 If X is a graph with n {1,,9} {4} vertices having symmetry group D n then H(X) = C(D n ) The list of such graphs is as follows (a) n-gons with n = 4 and their complements (b) 8-spoke wheel and its complement (c) 9-stars

19 T Banica / Journal of Functional Analysis 224 (2005) Proof Let X be as in the statement By relabeling vertices of the first 9-star we see that its complement is the second 9-star Thus the list of graphs is closed under complementation It is enough to show that X or X c appears in the list By replacing X X c we may assume that the valence k of vertices is smaller than (n 1)/2 Thus n 2k + 1 For k = 0 the possible graphs are the point, the 2 points and the 3 points These are the 1-gon, the complement of the 2-gon, and the complement of the 3-gon, all 3 in the list For k = 1wehaven 3 On the other hand the graph must be a union of segments, so the only solution is the 2 segments But here n = 4 For k = 2wehavethen-gons with n 5, all of them in the list For k = 3wehaven 7 The cases n = 7, 9 are excluded, because the incidency matrix must have 3n/2 values of 1 above the diagonal In the n = 8 case the graph corresponds to a polynomial of form Q(z) = z a +z 4 +z 8 a with a = 1, 2, 3 For a = 1 this is the 8-spoke wheel, in the list, for a = 2 we get the 2 tetrahedra, not dihedral, and for a = 3 we can relabel vertices and we get the 8-spoke wheel again For k = 4wehaven = 9 Consider the associated Q polynomial Q(z) = z a + z b + z 9 b + z 9 a Here a = b are from {1, 2, 3, 4} Since a,b are not both equal to 3, one of them, say a, is prime with 3, and by relabeling vertices 0,a,2a,,7a we can assume a = 1 For b = 2, 3 we get the 9-stars and for b = 4 we can relabel vertices and we get the first 9-star Theorem 42 applies to n-gons with n = 4 For the 8-spoke wheel we have Q(w k ) = w k + ( 1) k + w k and computation gives the distinct numbers 3, ±1 and 1 ± 2, so Theorem 42 applies as well Let X be a 9-star and assume that we have an equality of the form Q(w k ) = Q(w l ) with 0 k <l 4 The numbers Q(w n )/2 are sums of numbers cos(2nπ/9), and we see that the only equalities between two such sums are those of form x + x = x + x In particular we must have cos(2kπ/9) = cos(2lπ/9) But this is impossible because cos(2xπ/9) is decreasing on [0, 4] Thus Theorem 42 applies to both 9-stars and we are done 5 Tannaka Galois duality An arbitrary Hopf C -algebra with a faithful coaction v : C(X) C(X) H is quite an abstract object However, a useful description is obtained after classifying its irreducible corepresentations, together with their fusion rules In this section, we present a categorical and topological approach to this problem, by using Woronowicz s Tannakian duality [16] and the spin planar algebra P(X) constructed by Jones in [9,10] The main result will be a Tannaka Galois type correspondence between pairs (H, v) and subalgebras P P(X), somehow in the spirit of the correspondence found by Kodiyalam et al in [12]

20 262 T Banica / Journal of Functional Analysis 224 (2005) Recall from Section 1 that the main problem is to decompose tensor powers of v Consider the Hilbert space where the mth tensor power of v acts P m (X) = C(X) m In planar calculus both the input and the output are written in a 2 m matrix form, and the first thing to be done is to write elements of P m (X) in such a way Definition 51 Each m-fold tensor product of Dirac masses at points of X is identified with a 2 m matrix with coefficients in X, in the following way: δ i1 δ im = ( ) i1 i 1 i 2 i 2 i 3 i m i m i m 1 That is, we take the sequence of m points, we duplicate each entry, then we put it in the 2 m matrix, clockwise starting from top left We recall now some basic notions from Jones planar algebra formalism [9] A box is a rectangle in the plane It is convenient to assume that sides of the box are parallel to the real axis and imaginary axis We say that a box X is at left of a box Y if the horizontal sides of X and Y are on the same lines, and if X is at left of Y Same for on top A m-box is a box with 2m marked points, m of them on the lower side and m of them on the upper side If x P m (X) is a tensor product of Dirac masses, written in loop form, we can put indices on marked points in the obvious way We say that x is in the box Let m 1,,m k and n be positive integers Let T be a picture consisting of an output n-box, containing an input m i -box for each i, together with some non-crossing strings outside the input boxes These are strings connecting pairs of marked points, plus a finite number of closed strings, called circles Strings connecting input and output points are assumed to connect odd-numbered points to odd-numbered points and even to even, when numbering points on each k-box 1, 2,,2k clockwise starting from top left Such a picture, or rather its planar isotopy class, is called (m 1,,m k,n)-tangle, or just m-tangle Planar tangles act on tensors in the following way See Jones [9,10] Definition 52 Each (m 1,,m k,n)-tangle T defines a multilinear map P m1 (X) P mk (X) P n (X) in the following way If x 1,,x k and y are m 1,,m k -fold and n-fold tensor products of Dirac masses written in loop form, put each x i in the m i -input box of T and y in the output n-box of T Strings of T join now indices, and the number (x 1,,x k,y) T

21 T Banica / Journal of Functional Analysis 224 (2005) is defined to be 1 if all strings join pairs of equal indices and 0 if not Define T(x 1 x k ) = β c(t ) y (x 1,,x k,y) T y, where the sum is over all n-fold tensor products of Dirac masses, β is the number of elements of X and c(t ) is the number of closed circles of T The planar calculus for tensors is quite simple and does not really require diagrams It suffices to imagine that the way various indices appear, travel around and disappear is by following some obvious strings connecting them Some illustrating examples Example 51 Identity, multiplication, inclusion The identity 1 m is the (m, m)-tangle having vertical strings only The solutions of (x, y) 1 m = 1 are pairs of the form (x, x), so1 m acts by the identity ( ) ( ) j1 j 1 m j1 j m = m i 1 i m i 1 i m The multiplication M m is the (m, m, m)-tangle having 2 input boxes, one on top of the other, and vertical strings only It acts in the following way: (( ) ( )) ( ) j1 j M m l1 l m m l1 l = δ i 1 i m k 1 k j1 k 1 δ m jm k m m i 1 i m The formula xy = M m (x y) defines an associative multiplication of P m (X) The inclusion I m is the (m, m + 1)-tangle which looks like 1 m, but has one more vertical string, at right of the input box Given x written in loop form, solutions of (x, y) I m = 1 are elements y obtained from x by adding to the right a vector of the form ( ) l l ( ) j1 j I m m = i 1 i m l ( ) j1 j m l i 1 i m l This shows that I m is an inclusion of algebras, and that various I m are compatible with each other The inductive limit of the algebras P m (X) is a graded algebra, denoted P(X) Example 52 Expectation, Jones projection The expectation U m is the (m + 1,m)-tangle which looks like 1 m, but has one more string, connecting the extra 2 input points, that we suppose to be both at right of the

22 264 T Banica / Journal of Functional Analysis 224 (2005) input box ( ) ( ) j1 j U m j m+1 j1 j m = δ m i 1 i m i im+1 j m+1 m+1 i 1 i m This shows that U m is a bimodule morphism with respect to I m The Jones projection E m is a (0,m+ 2)-tangle, having no input box There are m vertical strings joining the first m upper points to the first m lower points, counting from left to right The remaining upper 2 points are connected by a semicircle, and the remaining lower 2 points are also connected by a semicircle We can apply E m to the unit of C E m (1) = ( ) i 1 i m j j i 1 i m k k The elements e m = (#X) 1 E m (1) are projections, and define a representation of the infinite Temperley Lieb algebra of index #X on the inductive limit algebra P(X) Example 53 Rotation The rotation R m is the (m, m)-tangle which looks like 1 m, but the first 2 input points are connected to the last 2 output points, and the same happens at right R 5 = The action of R m is best described in terms of Dirac masses R m ( δi1 δ im ) = δi2 δ im δ i1 Thus R m acts by an order m linear automorphism of P m (X), also called rotation Multiplications, inclusions, expectations, Jones projections, rotations generate in fact the set of all tangles, with the gluing operation described below Definition 53 Let T be a (m 1,,m k,n)-tangle and let S be a m i -tangle with m i among m 1,,m k The composition TS is obtained by superposing the output box of S on the input m i -box of T, after isotoping so that marked points match, then by removing the common boundary The colored planar operad P is the set of all tangles, with this gluing operation

23 T Banica / Journal of Functional Analysis 224 (2005) The composition U m I m consists of 1 m plus a floating circle, and by using the above formulae for actions of 1 m, I m and U m we get the following equality: (U m I m )(x) = U m (I m (x)) = (#X) x = 1 m (x) In general, composition of tangles corresponds to composition of maps We have a morphism from P to the colored operad of multilinear maps between spaces P(X), called action of P on P(X) This action commutes in some natural sense with the involution of P m (X) ( ) ( ) j1 j m i1 i = m i 1 i m j 1 j m This means that P(X) is a C -planar algebra, called spin planar algebra See [9] Definition 54 The graded linear space P(X) together with the action of P and with the involution is called spin planar algebra associated to X Let v : C(X) C(X) H be a coaction Consider the mth tensor power of v v m : P m (X) P m (X) H Computation using the magic biunitarity condition shows that each v m is a C - morphism See [4] for details Let P m be the fixed point algebra of v m P m ={x P m (X) v m (x) = x 1} Consider the rotation R m Rotating, then applying v m, then rotating backwards by Rm 1 is the same as applying v m, then rotating each m-fold product of coefficients of v Thus the elements obtained by rotating, then applying v m, or by applying v m, then rotating, differ by a sum of tensor products of Dirac masses tensor commutators in H v m R m (x) (R m id)v m (x) P m (x) [H,H] Let h be the Haar functional and consider the conditional expectation m = (id h)v m onto the fixed point algebra P m The square of the antipode being the identity, h is a trace, so it vanishes on commutators Thus R m commutes with m m R m = R m m The commutation relation n T = T m holds in fact for any (m, n)-tangle T These tangles are called annular, and proof in [4] is by verification on generators of the annular

24 266 T Banica / Journal of Functional Analysis 224 (2005) category In particular we get n T m = T m for any T, so the annular category is contained in the suboperad P P consisting of tangles T satisfying the following condition, where = ( m ) and i() is the number of input boxes T i(t ) = T i(t ) On the other hand multiplicativity of v m gives M m P Since P is generated by multiplications and annular tangles, it follows that P = P Thus for any tangle T the corresponding multilinear map between spaces P m (X) restricts to a multilinear map between spaces P m In other words, the action of P restricts to P and makes it a subalgebra of P(X) Definition 55 The sequence of spaces of fixed points of v m, together with the restriction of the action of P is called C -planar algebra associated to v Consider pairs (H, v) where v : C(X) C(X) H is a faithful coaction A morphism (H, v) (K, w) is a Hopf C -algebra morphism H K sending v ij w ij Isomorphism means morphisms in both senses, and this is the notion used in the statement below Theorem 51 IfQisaC -planar subalgebra of P(X) there is a unique pair (H, v) with v : C(X) C(X) H faithful coaction whose associated C -planar algebra is Q Proof This will follow by applying Woronowicz s Tannakian duality [16] to the annular category over Q This is constructed as follows Let n, m be positive integers To any element T n+m Q n+m we associate a linear map L nm (T n+m ) : P n (X) P m (X) in the following way: L nm T n+m : a n T n+m a n That is, we consider the planar (n, n + m, m)-tangle having an small input n-box, a big input n + m-box and an output m-box, with strings as on the picture of the right This defines a certain multilinear map P n (X) P n+m (X) P m (X) Nowwe put T n+m in the big input box What we get is a linear map P n (X) P m (X) This is called L nm The above picture corresponds to n = 1 and m = 2 This is illustrating whenever n m, suffices to imagine that in the general case all strings are multiple If n>mthere are n + m strings of a n which connect to the n + m lower strings of T n+m, and the remaining n m ones go to the upper right side and connect to the